@agent-pattern-labs/iso-orchestrator

@agent-pattern-labs/iso-orchestrator

Durable workflow primitives for AI-agent harnesses.

This package is for the layer above a single agent session. An agent runtime already knows how to think, call tools, and emit text. @agent-pattern-labs/iso-orchestrator adds the deterministic parts that should not live only in prompt prose:

• resumable, idempotent step() execution with file-backed records
• bounded parallel fan-out with forEach(..., { maxParallel })
• keyed mutexes so "same entity" work never runs concurrently
• worker heartbeats plus renewable leases for liveness / ownership tracking
• append-only-ish workflow events and durable state snapshots on local disk

It is intentionally generic. There is no built-in "dispatch a Codex worker" or "spawn an OpenCode task" primitive yet. Domain packages bring their own adapter for that part and use this package to enforce invariants around it.

Install

npm install @agent-pattern-labs/iso-orchestrator

What it stores

By default the package writes under .iso-orchestrator/ in the current working directory:

.iso-orchestrator/
  workflows/
    my-flow-<hash>.json
  locks/
    my-flow-<hash>/
      record.lock/
      mutex/
        company-role-<hash>.lock/

The workflow record contains:

• current workflow status (idle, running, completed, failed)
• durable JSON state
• step attempt counts and cached step results
• optional heartbeat snapshots and lease ownership records
• event history (workflow.running, step.started, step.completed, ...)

Quick example

import { runWorkflow } from '@agent-pattern-labs/iso-orchestrator';

const { value, record } = await runWorkflow(
  {
    workflowId: 'apply-batch-2026-04-25',
    dir: '.jobforge-runtime',
    initialState: { applied: 0, skipped: 0 },
  },
  async (workflow) => {
    await workflow.step('cleanup-geometra', async () => ({ ok: true }));

    const jobs = [
      { company: 'Anthropic', role: 'Staff Engineer' },
      { company: 'OpenAI', role: 'Member of Technical Staff' },
      { company: 'Anthropic', role: 'Staff Engineer' },
    ];

    const summary = await workflow.forEach(
      jobs,
      async (job) => {
        return workflow.step(
          `apply:${job.company}:${job.role}`,
          async () => {
            // Your own task-dispatch adapter goes here.
            return { outcome: 'APPLIED', company: job.company, role: job.role };
          },
          { idempotencyKey: `${job.company}:${job.role}` },
        );
      },
      {
        maxParallel: 2,
        mutexKey: (job) => `${job.company}:${job.role}`,
      },
    );

    await workflow.updateState((state) => ({
      ...state,
      applied: summary.fulfilled,
      skipped: summary.rejected,
    }));

    return summary;
  },
);

console.log(value.fulfilled, record.status);

API

runWorkflow(options, fn)

Creates or re-opens a workflow record, marks it running, executes fn, and marks the workflow completed or failed.

Options:

• workflowId: durable identifier for the logical workflow
• initialState: JSON value written only when the workflow record does not exist yet
• dir: optional storage root (defaults to .iso-orchestrator)
• now: optional clock injection for tests

openWorkflow(options)

Opens the workflow context without automatically running a top-level callback. Useful when you want finer control over lifecycle or only need inspection.

workflow.step(name, fn, options?)

Runs one load-bearing step and persists its result.

• If the same step already completed, the cached JSON result is returned.
• idempotencyKey lets one logical step name be reused safely across runs.
• retry can be a number (3) or { attempts, shouldRetry }.

Results must be JSON-serializable because they are persisted in the record.

workflow.withMutex(key, fn, options?)

Runs fn while holding a process-safe filesystem lock for key. Useful when two parallel tasks must never touch the same entity at once.

Options:

• timeoutMs: how long to wait for the lock
• pollMs: wait interval while the lock is held elsewhere
• staleAfterMs: optional stale-lock eviction threshold

workflow.heartbeat(key, detail?)

Persists the latest heartbeat for a named worker or phase. This is useful when you want external inspectors to tell whether a background task is still progressing even if it is not currently holding a mutex.

workflow.touchLease(key, { holder, ttlMs, detail? })

Acquires or renews a renewable lease for key.

• a missing, expired, or released lease is acquired
• the current holder can renew in place
• a different holder gets a WorkflowLeaseConflictError while the lease is active

Use this for "one worker owns this slot until its heartbeat expires" semantics without introducing a harness-specific queue.

workflow.releaseLease(key, holder?)

Marks a lease released. Passing holder is optional but recommended so a worker cannot accidentally release another worker's active lease.

workflow.forEach(items, fn, options?)

Bounded fan-out over a list.

• maxParallel controls concurrency
• mutexKey(item) optionally serializes related items
• stopOnError defaults to true

Returns a summary with results, fulfilled, and rejected.

workflow.updateState(updater)

Replaces the durable JSON state. Accepts either a full state object or a functional updater.

workflow.appendEvent(input) / workflow.getRecord()

Low-level helpers for custom lifecycle tracking and inspection.

Scope boundary

This package does not try to be Temporal.

What it does:

• local durable records
• load-bearing step caching
• process-safe local mutexes
• simple bounded concurrency

What it does not do yet:

• distributed queues
• remote workers
• harness-specific task dispatch APIs
• cron / scheduling

That narrower scope is deliberate. The goal is to let packages like JobForge stop expressing orchestration invariants only in prompt prose or Bash without forcing a heavyweight control plane.